Detect Wake Word (openWakeWord)

How it fits

Typical backends

Caveats

• CompatibilityConfigured wake words are either built-in model names (such as `alexa`, `hey_jarvis`, `hey_mycroft`) or absolute paths to a custom `.onnx` model; relative paths resolve against the bundled directory. The built-in weights are downloaded on first boot, so outbound HTTPS to the model host is required once, then cached.
• I/O contractAn empty input AudioFrame returns an empty list without invoking the models, multi-channel input is mixed to mono before resampling, and the output is unordered — sort downstream if a stable order is needed.
• Parameter interactionthreshold is a single hard cutoff applied uniformly to every loaded model; per-word thresholds require post-filtering downstream.
• Hard constraintinference_framework must be `onnx` or `tflite`; any other value aborts startup. `onnx` typically gives lower latency on x86 CPU, while `tflite` is the reference backend.
• LimitationInference is CPU-only regardless of host, and all configuration keys are captured once at startup, so runtime changes require a redeploy.